Research On Improved Algorithm Of High Precision And Real-time Image Matching Based On Video Seguence

In the displacement deformation monitoring,traditional monitoring methods have high measurement accuracy and wide application,such as level,total station,GPS and various contact sensors.However,due to the limitations of the surrounding environment,the defects of small scope,discontinuity and high cost are existed.Therefore,the application of visual measurement technology which realizes non-contact,low-cost and continuous large-scale photography has become a hot research issue in the field of deformation monitoring.At the same time,the development of image processing technology also provides a technical support for video sequence processing,making real-time dynamic deformation monitoring possible.An improved sFAST feature detection combined with BRIEF feature descriptive vector is purposed for feature points detection rapidly and subpixelated,and DBSCAN which is unsupervised learning method in machine learning is used to high-precision matching of feature points,and which is the basis to establish a set of dynamic deformation monitoring system for automatic extraction of vertical displacement and visualization of results.The important theoretical significance and engineering practical value for real-time and automatic high-precision deformation monitoring is proved by laboratory simulation test and field test.The main research contents in this paper summarized are as follows:?1?According to the local features of the image,only the interception ROI is processed in order to realize the rapid and real-time processing of video image,which can reduce the processing range and the amount of data greatly,on this basis,the advantages and disadvantages of the existing classical feature extraction algorithm are analyzed.The SIFT and SURF algorithms extract feature points with high accuracy and strong descriptor stability.However,the construction of image pyramid is complex,which greatly limits the extraction speed.At the same time,the high-dimensional feature description vector also reduces the matching speed in the later stage.Therefor,a fast and efficient and subpixelated algorithm combining sFAST and BRIEF is proposed on the basis of satisfying the detection accuracy.In terms of processing ROI,gaussian filtering is adopted to improve the noise sensitivity of the feature points of FAST extraction,and the feature detection results by FAST are interpolated to obtain sub-pixel coordinates,which effectively improves the accuracy of feature matching and displacement extraction.The principle is simple and is suitable for feature extraction of video sequences without large angle rotation and high intensity compression.Taking the extraction speed and the repetition rate of feature points as the evaluation criteria,the Gaussian parameter?was changed to simulate different fuzzy scenes caused by the distance,and the real image detection is carried out.The experiment shows that the algorithm is fast in operation and strong in extracting feature points,which has a certain robustness for fuzzy images.?2?For the binary vectors formed by BRIEF description algorithm,BF-Match based on hamming distance is carried out for rapid matching,but the fast speed is based on the sacrifice of accuracy.Therefore,DBSCAN belongs to unsupervised learning method is introduced in this paper to cluster the normalized two-dimensional point sample set composed of the distance and angle between the coarse matching result.According to when d_??is less than 3?acquired empirically?,the correct rate of BF-Match is above 50 percent.So the matching point pairs in the largest cluster are retained,so as to eliminate false matches.Different from the traditional improved algorithms,no threshold setting is required,and it is widely used.Finally,the correctness rate of the different density neighborhood parameters is calculated by the homography matrix to select optimal parameter solution,and the optimization and improvement effect is evaluated by comparing with the RANSAC results.Experiments show that the improved algorithm has a high precision and reliability.?3?Based on the image processing algorithm studied in this paper,the secondary development of the dynamic displacement deformation monitoring system is carried out on the single document MFC platform in the Visual Studio 2015 and Windows 7 system environment by installing OpenCV 3.2,TeeChart Pro Activex Control 2018 and List Control.The system includes the interface design and the function module implementation.The interface design consists of a menu bar and a work area divided into three areas,which is generous and easy to operate.The function module includes three parts:a image processing module,a displacement extraction module and a displacement visualization module.The whole system can realize real-time and high-precision automatic processing of video sequences,and provide more effective details for displacement deformation monitoring,and contribute to architectural design reference,disaster monitoring and early warning analysis.and enrich and develop the system of deformation monitoring.